A parallel stochastic method for solving linearly constrained concave global minimization problems (Q1200631)

A parallel stochastic algorithm is presented for the global minimization of a differentiable strictly concave function over a polytope. The approach is based on a multistart technique, which repeatly employs a global and a local phase during the solution process. In the global phase, a random search direction is selected and used to obtain a starting point for the local phase; the local phase then attempts to find a local minimum by starting from this point. A Bayesian stopping rule is used to terminate the algorithm with a high probability that all of the local minima have been found. Computational results are presented for more than 200 problems on a Cray X-MP EA/464 supercomputer. Only three classes of functions have been tested: quadratic, exponential and fixed charge functions (these are approximated by concave exponential functions in separable form).